Translation Isn't Enough: Multilingual AI Needs Storytellers
In Brazil, barraqueira is an insult. I have lived in Canada for 13 years, and I named my YouTube channel after it. A dictionary will tell you that a barraqueira is a woman who makes a scene. That is accurate, and it misses almost everything. She is the woman who says out loud what she thinks, tells uncomfortable truths, and gets called trouble for it. The word comes from barraco, a shack built without an architect, from whatever is at hand. Living in a second language, I came to see my English the same way: improvised, a little crooked, built without a plan, and still a home. The word mattered so much to me that it became the title of my master's thesis. Translate barraqueira as "troublemaker" and every word is correct. But the story disappears: the insult, the reclaiming, the shack that becomes a home.
Retelling, not translating
link:https://youtu.be/Fo7tFoEfbG4?si=_K5FaHO4qXscl02M
I write and make videos in Portuguese and English about language, culture and life between Brazil and Canada. Over the years I've learned one rule: my texts in different languages are never translations of each other. When I script a video for a Brazilian audience, I can lean on shared references, a certain rhythm of humour and a closeness that Brazilian Portuguese allows. When I tell the same story in English for Canadian viewers, I have to rebuild it. Some jokes need setting up. Some references need a sentence of context. Sometimes the whole point of the story moves, because what surprises a Canadian is not what surprises a Brazilian. The facts stay the same. The story is told again, for different people. This post is a small example. The video that introduces my channel opens in Portuguese with one line: "No Brasil, barraqueira é xingamento." For a Brazilian viewer, that sentence is a provocation, and it needs nothing else. For you, reading in English, it needed the whole opening of this post.
Why this matters for global companies Businesses have known for years that language decides whether people buy. In a 2020 survey of 8,709 consumers in 29 countries, CSA Research found that 76% prefer to buy products with information in their own language, and 40% will never buy from websites in other languages. Three in four said they were more likely to buy from a brand again if its customer care spoke their language. The detail I find most telling: people in countries with high English proficiency, such as Sweden and Germany, were among the most committed to content in their own language. They can read English. They simply trust a story told in their own. Now companies are handing much of that storytelling to AI. Support answers, product descriptions, onboarding emails and internal knowledge bases are increasingly drafted by models, often in dozens of languages at once. A model can produce grammatically perfect Portuguese and still sound like a stranger. It can choose a word that is correct in Lisbon and odd in São Paulo. It can be polite where a Brazilian customer expects warmth, or warm where a German client expects precision. None of these are translation errors. They are storytelling errors, and they cost trust.
What multilingual AI needs from storytellers The answer is not to stop using AI in other languages. It is to bring people who retell stories for a living into the loop. In my work evaluating AI systems, I have learned that an answer can be fluent and correct and still miss its reader. Closing that gap takes four things.
1.Judge meaning, not just accuracy. Checking that a sentence is correctly translated is the easy part. The harder question is whether it lands the way it was meant to for this reader, in this culture.
2.Write a voice guide for every language. One brand voice, written down separately for each language: how formal to be, what humour is allowed, which words to avoid. A model follows instructions better when someone has decided what "our voice in Brazilian Portuguese" actually sounds like.
3.Test with real phrasing from each place. People don't write to a company the way a textbook does. Prompts and evaluation sets should include slang, regional spellings and mixed-language messages, the way customers really type.
4.Give native speakers the final edit. For the content that matters most, such as launches, apologies and anything legal, a person who lives in that language should read it last.
Protecting the voice is editorial work.
The kaleidoscope
I once wrote a piece in English and Portuguese that I called a kaleidoscope. Turn it one way and you see one picture; turn it the other way and the same pieces make something new. That is what speaking two languages feels like to me, and it is what global companies are now asking their AI to do every day. Multilingual AI is one of the most useful tools a global company can have. But the model can only turn the kaleidoscope. Someone still has to know which picture each reader needs to see. That someone is a storyteller. Sometimes she is a barraqueira: the woman who raises her voice, even in a crooked second language, when the story is about to get lost.